Rajan G (@najarvg) header
Rajan G (@najarvg) profile picture

#cancerinformatics @uccancercenter. Science geek. Occasionally politics. Music, beer and burger devotee. Opinions my own! @najarvg@mastodon.online on #Mastodon

Chicago
Joined October 2013
Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)Photo by Rajan G (@najarvg)
We're proud to announce the launch of CS Learn, a new online learning platform that offers free access to select courses from Carnegie Mellon University’s top-ranked computer science program. Check out the course offerings: bit.ly/4C4770z
30
543
41
3,989
214,226
Rajan G retweeted
ShellGym is a good fit not only for practicing Linux 101. Here is an example of a Kubernetes gym, prepared by Kaloyan Preslavski, our community author 👏 It looks like a great format for drilling kubectl commands before a CKA/CKAD/CKS exam: labs.iximiuz.com/shell-gyms/…
6
22
258
9,526
Rajan G retweeted
the illusion of choice
32
104
15
1,543
56,594
The Sufis of Chettinad Tucked away in a quiet corner of Tamil Nadu lies one of the greatest clusters of palatial mansions in all of India. 🧵
7
112
7
753
30,958
There are so many flawed OCR benchmarks, and more come out every week. I gave a talk on how to evaluate them in @HamelHusain's course. Here are common issues, how to spot them, and a guide on evaluation 🧵
12
13
5
112
15,174
A 28-year-old UMass graduate student got a homework assignment to reproduce a famous economics paper, couldn't make the numbers match, and found a spreadsheet mistake in a study that politicians had used to argue for budget cuts. His name is Thomas Herndon. In 2010, two Harvard economists, Carmen Reinhart and Kenneth Rogoff, published a short paper called Growth in a Time of Debt. Its headline finding: once a country's public debt goes above 90% of its economy, growth falls off a cliff. In their numbers, average growth for that group was minus 0.1%. It came out right after the 2008 financial crisis, and it spread fast. In the US, the Paul Ryan budget cited it as its only evidence on what high debt does to growth. In Europe, an EU commissioner pointed to the research in a speech. In Britain, George Osborne cited the authors by name. Herndon was a PhD student at the University of Massachusetts Amherst, taking a class called Econ 753: Applied Econometrics. The big assignment was to pick a well-known paper and replicate it. He picked this one. He couldn't get the same results. He asked the authors for their spreadsheet, and they sent it. When he spotted the problem, he called his girlfriend over to make sure he wasn't seeing things. He wasn't. > One formula stopped five rows short > That quietly left out five countries: Australia, Austria, Belgium, Canada and Denmark > Some available years of data for other countries weren't included > The way the averages were weighted let a single bad year in one country count as much as decades in another Herndon and his professors, Michael Ash and Robert Pollin, published the critique in April 2013. Fixing the problems, average growth for the high-debt countries came out to 2.2%, not minus 0.1%. Reinhart and Rogoff publicly admitted the spreadsheet error. They disputed the other criticisms and defended their broader view. Economists still argue about debt and growth today. But the formula was wrong, and for about three years nobody had publicly caught it. A student doing homework did. Most people assume famous research has been checked many times. Often the first person to really check it is someone with nothing to lose and an assignment due. Have you ever found a mistake in your boss's spreadsheet and kept quiet?
85
559
66
4,787
896,657
Rajan G retweeted
I’ve been using Claude Code to make 3Blue1Brown-style explainer videos to teach myself textbooks and papers since the release of the Opus 5.5 model. While Opus 5.5 is great at creating animations, it’s not that good at teaching, so I’ve spent quite some time building a plugin to help it teach well. The result is quite astonishing. I've made almost a hundred videos and am really enjoying it. While building this plugin, I thought, instead of keeping it for my personal use, why not release it to everyone? So we merged everything into Heptabase’s official plugin. It’s free, open-source, and no sign-up is required. heptabase.com/plugin Once installed, you can give Claude a source (e.g., a PDF file) and ask it to create videos for you. Claude will read the source, plan a series of videos, and choose the best teaching method for each video. There are five teaching method in total, which I call “presenters”:
24
131
11
1,199
63,927
We added a new skill to an AI model by editing its weights directly. No training, 3 milliseconds, using a decision model without catastrophic memory loss- what MIT did without a SOTA LLM The teacher: vegaml, our decision model, fitted on 300 examples. Its decisions are like valleys in a landscape, and every answer rolls downhill into one, with honest confidence. We wrote those valleys straight into Qwen3.5-0.8B's weights, then removed the decision model. Qwen now knows on its own when to use tools and when it's ready to act: from 18% to 98% right Normally, that takes hours of training and risks breaking what the model already knows. kaggle.com/code/nandakishorc… embedding.in/blog/a-decision…
16
41
6
451
14,297
What happens when you combine embedding models with System 1 decision models? You get the intersection, which is VENN —> Vector Decision Models, a very cool OSS research project from Mark Russinovich that I was fortunate to review early and am excited to support.
24
59
7
552
83,788
Rajan G retweeted
As a vocal critic of the mayor’s budget, I thought it only fair to propose non-tax solutions. My op-ed today outlines a novel – and admittedly untested – idea I call a growth certificate: investors provide the city capital for a share of future tax growth above inflation. 1/16
7
5
1
48
6,509
Rajan G retweeted
How To Tell If Your Girlfriend Is A Spy:
749
7,382
1,009
65,195
6,239,901
New DDL project -> Help us crowdsource India's economic history! We're making a complete collection of India's district handbooks c.1951–2011, digitizing ALL the tables, harmonizing districts, variables, villages over time. You can help! 🧵👇 census.devdatalab.org/distri…
7
105
7
379
26,309
Rajan G retweeted
Stanford researchers gave an 8B language model 108× fewer document tokens. It became more accurate at verifying facts. The project is called Memorilla, an open-source memory system that lets LLMs use large collections of documents through just 16 learned memory tokens. Instead of repeatedly feeding an entire document collection into the context window, Memorilla encodes documents once and stores their embeddings. When a question arrives, a trainable memory module attends to the collection and produces 16 query-dependent vectors that the frozen LLM can read directly. The results are particularly interesting. On the FactKG benchmark, using the same Qwen3-8B model: → Full document context: 80.43% accuracy → Memorilla with 16 memory tokens: 84.33% accuracy → Document tokens passed to the LLM: 108× fewer Across eight benchmarks covering personalization, factual reasoning, and long-document QA, Memorilla achieved the highest average score among the evaluated methods. The approach isn't lossless. It can struggle with exact names and fine-grained facts that ordinary retrieval handles better. But the research demonstrates something worth studying: giving an LLM access to more text doesn't necessarily mean it will make better use of the information. The code, training recipes, and pretrained checkpoints are publicly available. Here's the GitHub Repo: github.com/snap-stanford/mem…
17
38
7
288
10,840
Read the fine print. This claims a $1M Phase 2 trial. What it actually contains is a 24 person Phase 2a repurposing an already approved drug that only lasted 2 weeks. Critically underpowered. The submissions are a mess, with a huge delay of registration over 100 days after study start, parallel vs crossover discrepancies, an unfavorable finding from ethics board, and notably a failure to report the results in January as was committed. They benchmark against trials much larger than theirs to make their numbers seem lower in a misleading way. The study uses an existing oral medicine, one virtual site, home-based monitoring, and a very small patient population. These are all important cost-reduction strategies, but none depends uniquely on generative AI. I’m supportive of AI reducing trial costs but this is wildly misleading and unhelpful.
7
13
1
105
7,217
My position on “late term abortions for convenience” is similar to my position on brain transplants. I don’t have one because they don’t exist.
483
10,092
409
75,420
When Surgery Center of Oklahoma opened our doors in 1997, our commitment to honest, transparent surgical pricing was seen as eccentric. Nearly three decades later, that approach has helped inspire a national movement toward greater transparency in healthcare. Today, hospitals are required to publish their prices, although patients still face plenty of obstacles when trying to determine what their care will actually cost. Meanwhile, we're still in the same building on Broadway Extension, still owned by the physicians who operate here, and still committed to giving patients one clear, all-inclusive price before their procedure. We've been publishing those prices online since 2009, long before anyone was required to do so. Nearly 30 years later, we still believe patients deserve to know what they're paying for. Price your procedure today: SurgeryCenterOK.com
8
31
4
182
7,641
Just published in @BloodJournal our @FredHutch investigator-initiated trial of CD20 CAR T in Waldenström macroglobulinemia—the largest reported CAR T series in WM. Developed in Brian Till’s lab; I had the privilege of serving as PI. The joy of bringing science from lab to clinic! ashpublications.org/blood/ar…… @IWMF_WM
2
14
3
46
3,477
China puts data centers in Malaysia to get access to compute they can't get in China due to sanctions. The US puts data centers in Malaysia because every state bans data centers. AGI happens. Malaysia nationalizes all the data centers and becomes the sole hegemon of the world.
JUST IN: Google to invest $15 billion in Malaysia data centers
184
1,653
135
21,738
596,433
Rajan G retweeted
Nanbeige4.3-33B claims to beat Qwen3.8-27B. on Hermes Agent tool calls on one 5090 it ties 82 vs 81, with 800 vs 799 on one-step calls, but finishes 25 of 30 real Hermes Agent tasks to Qwen's 29. tested: Nanbeige's 4-bit file (20.8 GB) against our Qwen3.8-27B rows on the same tests, one RTX 5090, llama.cpp. TL;DR: - tool calls over several turns: 82 vs 81 of 100, a tie - one-step tool calls: 800 vs 799 of 906, a tie - 30 Hermes Agent tasks at 64k context: 25 vs 29. Nanbeige only fits 64k with a compressed (q8_0) KV cache, Qwen ran without one - GPQA Diamond, thinking on: 81.3 vs 79.3, a tie on 198 questions - AIME 2026 at a 32k budget: 26 vs 25 of 30, a tie - five-task board, thinking off: 87.9 vs 93.2, Qwen ahead - speed: 71.7 vs 78.9 tok/s - the 6-bit file (29.0 GB): 80 of 100 over several turns vs 82 for the 4-bit, a tie, and its context tops out at 41k on 32 GB the card's scores come from other benchmarks at the highest thinking effort with no budget cap. these are our tests, so they check the claim on Hermes-style work, not the card's own numbers. on a 32 GB card, Qwen3.8-27B is still the pick for Hermes Agent. it matches Nanbeige on tool calls, finishes more tasks, runs faster and reaches 64k without compressing the cache. Nanbeige is worth a try if maths and science questions are your main use, where it ties Qwen. full report: github.com/notwitcheer/llm-b…
14
6
1
67
3,611
Rajan G retweeted
The era of System One Vision Model is here! Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI. Most embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For example, a bounding box is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go. VFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. What’s surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the chance to backspace. Everything that moves will be autonomous, and everything autonomous will need to think on its own chip. The next inference revolution is physical reasoning at the speed of reflex. And as always, it’s open paper, open source! Link in thread:
142
133
42
1,328
73,730